Xinlan Deng , Youhan Deng , Liang Qin , Weiwei Yao , Min He , Kaipei Liu
{"title":"Online evaluation method for MMC submodule capacitor aging based on CapAgingNet","authors":"Xinlan Deng , Youhan Deng , Liang Qin , Weiwei Yao , Min He , Kaipei Liu","doi":"10.1016/j.gloei.2025.03.002","DOIUrl":null,"url":null,"abstract":"<div><div>Submodule capacitor aging poses significant challenges to the safe operation of modular multilevel converter (MMC) systems. Traditional detection methods rely predominantly on offline tests, lacking real-time evaluation capabilities. Moreover, existing online approaches require additional sampling channels, thereby increasing system complexity and costs. To address these issues, this paper proposes an online evaluation method for submodule capacitor aging based on CapAgingNet. Initially, an MMC system simulation platform is developed to examine the effects of submodule capacitor aging on system operational characteristics and to create a dataset of submodule capacitor switching states. Subsequently, the CapAgingNet model is introduced, incorporating key technical modules to enhance performance: the Deep Stem module, which extracts larger receptive fields through multiple convolution layers and mitigates the impact of data sparsity in capacitor aging on feature extraction; the efficient channel attention (ECA) module, utilizing one-dimensional convolution for dynamic weighting to adjust the importance of each channel, thereby enhancing the ability of the model to process high-dimensional features in capacitor aging data; and the multiscale feature fusion (MSF) module, which integrates capacitor aging information across different scales by combining fine-grained and coarse-grained features, thus improving the capacity of the model to capture high-frequency variation characteristics. The experimental results reveal that the CapAgingNet model achieves a TOP-1 accuracy of 95.32 % and a macro-averaged F<sub>1</sub> score of 95.49 % on the test set, thereby providing effective technical support for online monitoring of submodule capacitor aging.</div></div>","PeriodicalId":36174,"journal":{"name":"Global Energy Interconnection","volume":"8 3","pages":"Pages 420-432"},"PeriodicalIF":1.9000,"publicationDate":"2025-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Global Energy Interconnection","FirstCategoryId":"1087","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2096511725000477","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"ENERGY & FUELS","Score":null,"Total":0}
引用次数: 0
Abstract
Submodule capacitor aging poses significant challenges to the safe operation of modular multilevel converter (MMC) systems. Traditional detection methods rely predominantly on offline tests, lacking real-time evaluation capabilities. Moreover, existing online approaches require additional sampling channels, thereby increasing system complexity and costs. To address these issues, this paper proposes an online evaluation method for submodule capacitor aging based on CapAgingNet. Initially, an MMC system simulation platform is developed to examine the effects of submodule capacitor aging on system operational characteristics and to create a dataset of submodule capacitor switching states. Subsequently, the CapAgingNet model is introduced, incorporating key technical modules to enhance performance: the Deep Stem module, which extracts larger receptive fields through multiple convolution layers and mitigates the impact of data sparsity in capacitor aging on feature extraction; the efficient channel attention (ECA) module, utilizing one-dimensional convolution for dynamic weighting to adjust the importance of each channel, thereby enhancing the ability of the model to process high-dimensional features in capacitor aging data; and the multiscale feature fusion (MSF) module, which integrates capacitor aging information across different scales by combining fine-grained and coarse-grained features, thus improving the capacity of the model to capture high-frequency variation characteristics. The experimental results reveal that the CapAgingNet model achieves a TOP-1 accuracy of 95.32 % and a macro-averaged F1 score of 95.49 % on the test set, thereby providing effective technical support for online monitoring of submodule capacitor aging.